I am the author of the AddGrain Avisynth filter, still on my web page. Though I no longer remember whether that was also the basis Milan used to add to ffdshow.
So while I don't like high grain film I obviously do think grain has some value.
What I believe is that the human visual system can easily notice an image that has no energy at the higher spatial frequencies. That is, no detail, plastic faces, excessive smoothness. This can be caused by blurry telecine practices, excessive filtering, repeated filter processing, over-compression, or simply by upscaling an image from some lower resolution as done with TV upconverts.
Either way the softness sticks out and can seem annoyingly artificial even if it is otherwise a very clean image.
Enter grain. It is mostly high frequency random noise. I'm going to go out on a limb here and speculate the human visual system has evolved to see meaningful images even in the presence of rain, dust, fog, tree branches, etc. So with some grain we may make allowances for noise in the image instead of just seeing an overly soft picture. In other words, we may think the picture is fine even though we may have some trouble seeing it.
This can probably be useful for hiding the fact that something was filtered or upconverted somewhat.
However totally random grain can also have course detail that obscures some of the lower frequency visual information that we could actually see and use. Meanwhile grain is fairly expensive to encode. For this reason I'm a fan of FGT-like technology where we just add the random noise at the time of decoding (AFTER possible upscaling). And most recently I've been considering ways to only add hi-pass filtered semi-random noise/grain that does not interfered with the lower frequencies of the image that were actually captured and encoded properly.
Just some thoughts.